How To Use Keras Trained CNN Models

Introduction

Keras is a popular deep learning api. It can run on top of Tensorflow, CNTK and Theano frameworks. Keras provides an easy to use interface which makes deep learning practice straight forward. It is widely used thus resources are easily accessible.

Objective

This article aims to give an introductory information about using a Keras trained CNN model for inference. This article does not contain information about CNN training.

Audience

This article assumes introductory information about python and Convolutional Neural Networks. For those who lack information may first begin with information from following resources.

• For python use Python For Beginners
• For Convolutional Neural Networks use CS231n Convolutional Neural Networks for Visual Recognition

Software Installation

Keras is a high level API. It requires a back-end framework to be installed. In this article, Tensorflow is used. Keras can transparently select CPU or GPU for processing. If use of GPU is desired, assuming presence of a  proper graphics card with a decent GPU, relevant drivers needs to be installed.

Installation is not a simple procedure. Prepare a Ubuntu System for Deep Learning can be read for installation details.

Trained Models

Training a CNN model requires specialization, a lot of data and decent hardware. Transfer learning may simplify those requirements but it is not in the scope of this article.

Keras provides already trained models. Trained models and information about how to use them can be found in Keras Applications. Those models are trained using Imagenet dataset.

Additional models can be found in my GitHub page which are created as part of my emotion recognition study. Model files can be found at deep-emotion-recognition repository. Those models are trained using FER-13 dataset which contains 7 emotions. Rest of the article uses emotion recognition models from my GitHub page.

Application Code 

Processing Pipeline

	from keras.preprocessing.image import ImageDataGenerator
	from keras.models import load_model
	VAL_BATCH_SIZE = 1
	def get_validation_generator(data_dir,size):
	test_datagen = ImageDataGenerator(rescale=1. / 255)
	# batch_size=1 and shuffle=False; we want validation data be exactly same,
	# validation accuracy exactly same for the same model and same data
	validation_generator = test_datagen.flow_from_directory(
	data_dir + 'Val/',
	target_size=size,
	batch_size=VAL_BATCH_SIZE,
	classes=None,
	class_mode='categorical', shuffle=False)
	return validation_generator
	def test(model_file, data_dir, dimension, prefix):
	size = (dimension, dimension)
	validation_generator = get_validation_generator(data_dir,size)
	VALID_DATA_COUNT=len(validation_generator.classes)
	print("Loading saved model: %s." % model_file)
	# load and compile the model
	model=load_model(model_file)
	#model.compile(optimizer='sgd',
	# loss='categorical_crossentropy',
	# metrics=['accuracy']) if additioanal metrics are required un comment and add here
	validation_result = model.evaluate_generator(validation_generator, VALID_DATA_COUNT / 1) # validation batch size = 1
	print(prefix)
	print(validation_result)
	def main():
	# this model will be loaded
	MODEL_FILE = '/mnt/sda2/dev_root/work2.1/merged/merged_models/1522678107.24-10033137-7-fer-1024-fear+sad/checkpoints/w.001-0.6818-1.68.hdf5'
	# images will be resized to this dimension
	DIMENSION = 224
	# for each dataset added one prefix
	prefixes=['fer_aligned']
	data=['/mnt/sda2/dev_root/dataset/original/fer2013_224_NEW_CLEAN/']
	for pref, data_dir in zip(prefixes,data):
	test(MODEL_FILE, data_dir, DIMENSION,prefix)
	if __name__ == '__main__':
	main()

view raw test_image.py hosted with ❤ by GitHub

This code is pretty straight forward. For loading a modal a load_model utility method is used. For loading images image generator provided by Keras is used. Please not that 1 is used as batch size. This is because for some reason using batch sizes other than 1 resulted in slightly different validation results for the same model at consecutive executions which is not acceptable.

For dataset either original dataset can be downloaded from original Kaggle page or from repository under dataset directory. Also note that the application only uses images found under Val directory.

Additional emotion datasets can be used. Some example datasets are:

• SFEW dataset which can be obtained upon request. For details see https://cs.anu.edu.au/few/AFEW.html.
• CK/CK+ dataset which can be obtained freely after accepting terms of use. For details see http://www.consortium.ri.cmu.edu/ckagree/.

Conclusion

We successfully loaded and evaluated a trained CNN model using Keras library. For full code listings you may check my GitHub source code repository.